Overview

bhenrym14/mistral-7b-platypus-fp16 is a 7 billion parameter instruction-tuned model built upon the Mistral-7B-v0.1 base architecture. It was fine-tuned by bhenrym14 using the Open-Platypus dataset, a collection designed to improve instruction-following abilities. The fine-tuning process involved a QLoRA method with a rank of 64.

Key Capabilities & Performance

This model demonstrates solid performance across several benchmarks:

• ARC (25 shot): 62.80
• Hellaswag (10 shot): 84.12
• MMLU (5 shot): 64.20

While its perplexity performance at shorter context lengths (512 tokens) is strong, the README notes that its competitiveness with larger models or those with advanced context extension techniques may be impacted at longer contexts, potentially due to its sliding window attention mechanism or model size.

Usage Notes

Users should be aware that the Mistral architecture currently requires installing transformers from source. The model was trained with a legacy Airoboros system prompt (version <2.0), and details for prompting can be found in the model card for bhenrym14/airoboros-33b-gpt4-1.4.1-lxctx-PI-16384-fp16.